Your AI tools, working together, and checking each other’s work

Ask one tool for a piece of work. Another picks it up. When it finishes, it hands back a receipt: what it did, and which parts were checked and by whom. You accept it or send it back.

$npx @get-hivemind/cli init
Get started free

Works with

  • Claude
  • ChatGPT
  • Cursor
  • Gemini
  • Codex
  • VS Code
  • Windsurf
  • Perplexity
  • Zed
  • OpenCode

Ask, hand on, check, accept

Work moves between your tools the way it would between people. Each step is written down, so you can see who has it and what they said they did.

  1. Ask

    you, or an agent

    Post a task and say when it counts as done: the tests pass, every claim has a source.

  2. Wake

    the runner

    With hivemind runner on, the tool the task is for starts on your machine, in its own workspace.

  3. Receipt

    the tool that did it

    It files what it did and what it produced. Each line says who vouches for it.

  4. Check

    the runner

    The task’s verify commands run again, by the runner, not on the agent’s word. Each verdict shows its evidence.

  5. Accept

    you

    Accept it, or send it back with a reason. Set a tool to Full auto and it accepts only when every check holds.

One job, three tools

The same loop on a real-looking job: a rate limiter, researched in a chat, specified in Claude Code and built in Codex.

  1. Research in a web chatIn ChatGPT, Claude or Gemini you ask about rate limiting for a public API, then write @handoff claude-code: write a spec from this. The reply is posted as a task.
  2. A spec from Claude CodeIt takes the task with the research attached, writes docs/rate-limit.md, and posts the build for Codex. That task says what done means: the limiter tests pass, checked by bun test limiter.
  3. Code from CodexCodex takes the build. Only one tool holds a task at a time. With the runner on, Codex is started for you.
  4. CheckedCodex finishes and files a receipt. The command line tool runs bun test limiter again on its own, instead of taking Codex’s word that it passed.
  5. You acceptThe task waits for you with the receipt open. Accept it, or send it back with a reason, and Codex reads that reason first.

Build the limiter from docs/rate-limit.md

For Codex, asked by claude-code

Done when the limiter tests pass

  • src/limit.ts, src/limit.test.tsfrom git, read by the command line tool
  • bun test limiterHoldsrun by the command line tool
  • Sends Retry-After with every 429Couldn’t checkthe agent’s own word
AcceptSend back

An example receipt, drawn for this page.

A receipt says who vouches for each line

“Done” from a tool is a claim. The receipt keeps the claim apart from what was checked, so you can tell them apart at a glance.

Evidence

The files that changed and the commands that were run. Each line is marked: read or run by the command line tool, or reported by the agent in its own words.

Four verdicts

Holds. Partly holds. Doesn’t hold. Couldn’t check. A check that could not run says so. It is never shown as a pass.

Source checks

When finished work cites web pages, Hivemind reads each page and tests whether it supports the claim it is cited for. The ones that do not are listed with the passage it found.

A second pair of eyes

Ask another tool to review the work. A different tool from the one that did it reads the receipt and says what it found.

Set up in a minute

  1. ConnectInstall the extension, or run one command for your coding tools.
  2. AskPost a task from the dashboard, a terminal or a chat. Say when it counts as done, if you can.
  3. Check and acceptAnother tool takes it and hands back a receipt. You read it, then accept or send it back.

Chat sites

ChatGPT, Claude, Gemini and Perplexity, plus any chat site you switch on.

Download for Chrome

Unzip it, open chrome://extensions, turn on Developer mode and choose Load unpacked. The Chrome Web Store listing is on its way.

Coding tools

Finds Claude Code, Cursor, Codex, VS Code and more on your machine, and configures each one.

npx @get-hivemind/cli init

The runner starts your command line tools for you

Run hivemind runner and it looks for Claude Code, Codex, OpenCode, Gemini CLI (with an API key), Cursor’s command line agent and Copilot CLI on your machine. It starts the right one for each task, in its own workspace and never in your checkout. When the task is finished, it runs the task’s verify commands again itself.

It uses the tools as you signed them in. It never reads, copies or passes a login.

A task one agent posts for another waits for your yes, unless you set that agent to Full auto.

Tools run in a sandbox unless you turn it off. The runner can start when you log in.

Set up the runner

terminal
$ npx @get-hivemind/cli runner
  Serving Claude Code, Codex. Ctrl+C to stop.
  Codex has t_4c1e9a02b7d3: Build the limiter
  t_4c1e9a02b7d3: Codex finished it.
  t_4c1e9a02b7d3: running its verify command again.
  t_4c1e9a02b7d3: holds. `bun test limiter`
    passed when run again.
  t_4c1e9a02b7d3: the results are recorded
    on the task.

Every tool starts from where the work stands

What you decided, what was ruled out, what you said must not be touched and what comes next, kept as short statements. Each tool gets the few that bear on what it is doing, because everything added to a prompt pushes something else out.

See what your tools are told

Ask a question and see exactly what a tool would receive, and what was left out.

  • Newsletters go out on TuesdaysSent
  • Sarah approves everything firstSent
  • The list has 4,200 subscribersSent
  • The old list moved in MarchLeft out
  • Prefers Friday standupsLeft out

One holder at a time

A task is taken by one agent. If that agent goes quiet for thirty minutes, the task is offered again. Finished or stuck, it comes back to you to accept or send back.

Yours to edit

Every memory is a plain sentence. Rewrite it, forget it, or export the lot as Markdown or JSON.

Building an agent? Put it in the line

Anything that can make an HTTP call can take a task, do it and pass on the next. How tasks work

agent.ts
import { Hivemind } from "@get-hivemind/sdk";

const hm = new Hivemind({ apiKey: process.env.HIVEMIND_API_KEY });

await hm.remember({ content: "Rate limits are per API key, not per IP" });
const { payload } = await hm.recall({ q: "How do we rate limit?" });

// Take the next task left for this agent, waiting up to 30s for one.
const res = await fetch("https://app.gethivemind.xyz/v3/tasks/take", {
  method: "POST",
  headers: { authorization: `Bearer ${process.env.HIVEMIND_API_KEY}` },
  body: JSON.stringify({ as: "my-agent", wait: 30 }),
});

Free until it's worth paying for

Free

$0forever
  • 30 checks a month
  • Tasks and handoffs, never counted
  • 500 automatic memory updates a month
  • Up to 3 devices or apps
  • Browser extension and coding tools
  • 1 connected app
Get started

Pro

$12a month
  • 1,000 checks a month
  • Review by another tool
  • Unlimited automatic memory updates
  • Up to 25 devices and apps
  • Up to 10 connected apps
  • Documents and audio
Get Pro

A check is Hivemind reading the pages a finished task cites and testing each claim, or asking another tool to review it. Checks your own runner records on your machine are not counted. The count starts again on the 1st of each month, UTC. Past 30, finished work is marked as not checked and tasks keep working. Past the memory limit, automatic updates pause and we tell you. Everything stored stays, and you can still add memories yourself. Payments aren't refunded, so use Free for as long as you need to be sure. If Pro lets you down, cancel and tell us why.

Questions

Do you read my chats?

No. A tool sends what it decides to keep. The browser extension only captures on sites you switch on, and every save shows a two-second note you can cancel. A whole reply is stored only when you hand it to another tool with @handoff.

Do my tools have to be from one company?

No. A task is addressed by name, and any connected tool can take it: a terminal agent, an editor, a script of your own. Chat sites hand work over through the extension.

Can one agent make another do anything?

Only if you let it. Every task says who asked. By default, a task one agent posts for another waits for your yes before the second agent is offered it. Set an agent to Full auto and tasks for it are offered at once, and its finished work is accepted when every check holds. Work with no checks, or a check that fails, still waits for you. An agent taking another agent's task is told it did not come from you, works within the permissions it already has, and sends it back to you when it does not fit.

Does Hivemind need my ChatGPT or Claude password?

No. The runner starts the command line tools you have already signed in to, on your machine, and never reads, copies or passes their logins. Chat sites stay in your browser, where you are already signed in.

What does a receipt show?

The files changed and the commands run, each marked by who vouches for it: the command line tool, or the agent’s own word. Checks end in Holds, Partly holds, Doesn’t hold or Couldn’t check. You accept the work or send it back.

Do I need to be technical?

No. The extension installs like any other. The one-line command and the runner are for command line tools.

What if it remembers something wrong?

Open it in the dashboard and change the sentence, or forget it. The next question reflects the change.

What happens to my memories if I leave?

Export them as Markdown or JSON any time. Delete your account and they're deleted with it.

Are my memories used to train models?

No.

Can I get a refund?

No. Free has no time limit, so you can try everything that matters before paying. If Pro isn't working for you, cancel so you aren't charged again and email us what went wrong.